EoRFlow

Simulation-based inference for 21cm reionization signals: a set of 21cm summary statistics paired with neural density estimators that infer astrophysical parameters or the neutral-fraction history from a lightcone.

Summaries

• spherically (1D) and cylindrically (2D) averaged power spectra
• CNN-learned summary from 21cm lightcones
• transformer-learned self-supervised summary, precomputed with SKATR

Inference models

• conditional normalizing flow (affine coupling, FrEIA)
• conditional flow matching (CFM)

Layout

src/models/        density estimators (flow, CFM) and the CNN summary
src/training/      training entry point
src/evaluation/    posterior sampling, metrics, plotting
src/data_tools/    dataset loaders, power spectra, noise injection
config/            configuration templates
scripts/           split generation, config generation, cluster submission

Getting started

Training and evaluation are two separate passes. Training writes a checkpoint; it does not write posterior samples.

# 1. edit the templates -- every path in them is a <placeholder>
cp config/train.yaml config/my_train.yaml

# 2. train
python src/training/train.py --config config/my_train.yaml

# 3. sample posteriors from the trained model
python src/evaluation/sample.py --config config/my_eval.yaml

Outputs are written to {project_root}/{output_subpath}/{mode}/{data}/{output_tag}/.

Splits

When several models must be compared, they have to share one held-out test set, otherwise the comparison leaks. scripts/generate_loreli_split.py writes an explicit split.npz of train/val/test indices, and scripts/generate_loreli_budget_splits.py derives nested per-budget subsets from it for data-scaling studies. Point split_indices_path at the result rather than relying on test_fraction.

Note that sample.py writes samples.npz whose row order does not follow test_indices.npy. Rows are aligned with the labels stored in the same file, so per-run metrics are correct, but joining two runs by row index compares different simulations. Match on the label vector instead.

Note on paths

This code was developed on a specific HPC system and the scripts under scripts/ contain absolute paths and SLURM directives from that environment. They are included as working examples of the full pipeline rather than as portable tools; adapt the paths before use.

Publications

EoRFlow was introduced in, and has been used for, the following work:

• Direct reconstruction of the reionization history from 21cm 2D power spectra
  Y. Pietschke, C. Heneka, T. Schlenker, A. Ore, B. Schosser — JCAP (2025)
  The original EoRFlow inference pipeline.
  Code: paper · journal · arXiv

• Constraining reionization morphology and source properties with 21cm galaxy cross-correlation surveys
  Y. Pietschke, A. Hutter, C. Heneka — A&A (2026)
  Application to inference from 21cm-galaxy cross-power spectra.
  journal · arXiv

• Cross-simulator transfer with foundation model summaries: Towards robust SKA-era reionization inference
  Y. Pietschke, C. Heneka, A. Ore, R. Meriot — submitted to A&A (2026)
  Frozen self-supervised SKATR summaries transferred across simulators and to SKA-AA* noise, with a conditional flow matching inference head.
  Code: main · arXiv

• SKA Science Data Challenge 3b
  EoRFlow contribution to the SDC3b inference challenge.
  Code: SKA-challenge · results